Here is design doc:
https://docs.google.com/document/d/1yIAYmbvL3JxOKOjuCyon7JhW4cSv1wy5hC0ApeGMV9s/pub

Here is the implementation:
https://golang.org/cl/12544043/

I've benchmarked the change of a real app where chans were abandoned in favor of other
sync primitives due to perf reasons:
https://groups.google.com/d/msg/golang-dev/0IElw_BbTrk/cGHMdNoHGQEJ
and the change provides 23% end-to-end speedup making chans significantly faster than
the other synchronization.